Designing Random Allocation Mechanisms: Discussion …Two Well-Studied Ordinal Mechanisms Coincide...

Designing Random Allocation Mechanisms:Discussion

Marek Pycia

UCLA

February 19, 2011

Key Assumptions about the Environment

� Transfers cannot be used

� Resale can be fully controlled

� Many agents

� Agents should be treated symmetrically – hence reliance onrandomization

Key Assumptions about the Environment

� Transfers cannot be used

� Resale can be fully controlled

� Many agents

� Agents should be treated symmetrically – hence reliance onrandomization

Key Assumptions about the Environment

� Transfers cannot be used

� Resale can be fully controlled

� Many agents

� Agents should be treated symmetrically – hence reliance onrandomization

Key Assumptions about the Environment

� Transfers cannot be used

� Resale can be fully controlled

� Many agents

� Agents should be treated symmetrically – hence reliance onrandomization

Good Ordinal Mechanisms

� Strategy-proof

� Ordinally efficient

� Symmetric (equal treatment of equals)

Two Well-Studied Ordinal MechanismsCoincide in the Limit!

� Random Priority (Abdulkadiroglu and Sonmez 1999)

� Probabilistic Serial (Bogomolnaia and Moulin 2001)

Che and Kojima 2010:

� Random Priority and Probabilistic Serial (i) converge as wereplicate the economy and (ii) coincide in the limit of amarket with continuum of agents

� In the continuum economy, they are strategy-proof, ordinallyefficient, and symmetric

Are there any other mechanisms with these good properties?

Core from Random Endowments (Top Trading Cycles) coincidewith Random Priority already in small markets (Abdulkadirogluand Sonmez 1999, Pathak and Sethuraman 2011)

Two Well-Studied Ordinal MechanismsCoincide in the Limit!

� Random Priority (Abdulkadiroglu and Sonmez 1999)

� Probabilistic Serial (Bogomolnaia and Moulin 2001)

Che and Kojima 2010:

� Random Priority and Probabilistic Serial (i) converge as wereplicate the economy and (ii) coincide in the limit of amarket with continuum of agents

� In the continuum economy, they are strategy-proof, ordinallyefficient, and symmetric

Are there any other mechanisms with these good properties?

Core from Random Endowments (Top Trading Cycles) coincidewith Random Priority already in small markets (Abdulkadirogluand Sonmez 1999, Pathak and Sethuraman 2011)

Two Well-Studied Ordinal MechanismsCoincide in the Limit!

� Random Priority (Abdulkadiroglu and Sonmez 1999)

� Probabilistic Serial (Bogomolnaia and Moulin 2001)

Che and Kojima 2010:

� Random Priority and Probabilistic Serial (i) converge as wereplicate the economy and (ii) coincide in the limit of amarket with continuum of agents

� In the continuum economy, they are strategy-proof, ordinallyefficient, and symmetric

Are there any other mechanisms with these good properties?

Core from Random Endowments (Top Trading Cycles) coincidewith Random Priority already in small markets (Abdulkadirogluand Sonmez 1999, Pathak and Sethuraman 2011)

Two Well-Studied Ordinal MechanismsCoincide in the Limit!

� Random Priority (Abdulkadiroglu and Sonmez 1999)

� Probabilistic Serial (Bogomolnaia and Moulin 2001)

Che and Kojima 2010:

� Random Priority and Probabilistic Serial (i) converge as wereplicate the economy and (ii) coincide in the limit of amarket with continuum of agents

� In the continuum economy, they are strategy-proof, ordinallyefficient, and symmetric

Are there any other mechanisms with these good properties?

Core from Random Endowments (Top Trading Cycles) coincidewith Random Priority already in small markets (Abdulkadirogluand Sonmez 1999, Pathak and Sethuraman 2011)

Asymptotically Only One Good OrdinalMechanism

Theorem ( Liu and Pycia 2011). In the continuum economy, everymechanism that is strategy-proof, ordinally efficient, andsymmetric coincides with Random Priority/Probabilistic Serial foralmost every preference distribution.

If we restrict attention to ”continuous” mechanisms then:

� In the continuum economy, Random Priority/ProbabilisticSerial is the only strategy-proof, ordinally efficient, andsymmetric mechanism.

� If a sequence of symmetric and strategy-proof mechanisms isasymptotically OE as we replicate the economy, then thesequence converges to Random Priority/Probabilistic Serial.

Asymptotically Only One Good OrdinalMechanism

Theorem ( Liu and Pycia 2011). In the continuum economy, everymechanism that is strategy-proof, ordinally efficient, andsymmetric coincides with Random Priority/Probabilistic Serial foralmost every preference distribution.

If we restrict attention to ”continuous” mechanisms then:

� In the continuum economy, Random Priority/ProbabilisticSerial is the only strategy-proof, ordinally efficient, andsymmetric mechanism.

� If a sequence of symmetric and strategy-proof mechanisms isasymptotically OE as we replicate the economy, then thesequence converges to Random Priority/Probabilistic Serial.

Asymptotically Only One Good OrdinalMechanism

Theorem ( Liu and Pycia 2011). In the continuum economy, everymechanism that is strategy-proof, ordinally efficient, andsymmetric coincides with Random Priority/Probabilistic Serial foralmost every preference distribution.

If we restrict attention to ”continuous” mechanisms then:

� In the continuum economy, Random Priority/ProbabilisticSerial is the only strategy-proof, ordinally efficient, andsymmetric mechanism.

� If a sequence of symmetric and strategy-proof mechanisms isasymptotically OE as we replicate the economy, then thesequence converges to Random Priority/Probabilistic Serial.

Intensive vs Extensive Margins

Manea 2009: Random Priority does not become ordinally efficientin the limit as the number of object types grows.

Ordinal and Cardinal Mechanisms

� In large markets with single-unit demands Random Priority isstrategy-proof, ordinarily efficient, symmetric, simple, anddoes not pose complex strategic problems for participants

� Cardinal mechanisms offer the premise of eliciting preferenceintensity, and hence implement Pareto-efficient outcomes

Ordinal and Cardinal Mechanisms

� In large markets with single-unit demands Random Priority isstrategy-proof, ordinarily efficient, symmetric, simple, anddoes not pose complex strategic problems for participants

� Cardinal mechanisms offer the premise of eliciting preferenceintensity, and hence implement Pareto-efficient outcomes

Trade-offs In Designing Cardinal Mechanisms

� Pseudo-market of Hylland and Zeckahuser 1979 :strategy-proof and efficient in the limit; challenge – how todecompose a random allocation?

� A-CEEI of Budish 2010: excellent with multi-unit demand,with single-unit demand becomes effectively ordinal (reducesto Random Priority)

� CADA of Abdulkadiroglu, Che, and Yasuda : a nicecompromise between strategy-proofness and accounting forpreference intensity

� Boston mechanism: good efficiency properties in equilibrium(Abdulkadiroglu, Che, and Yasuda 2011) but failsstrategy-proofness even in the limit (Abdulkadiroglu andSonmez 2003, Kojima and Pathak 2009) and has badredistributive properties (Pathak and Sonmez 2008)

Trade-offs In Designing Cardinal Mechanisms

� Pseudo-market of Hylland and Zeckahuser 1979 :strategy-proof and efficient in the limit; challenge – how todecompose a random allocation?

� A-CEEI of Budish 2010: excellent with multi-unit demand,with single-unit demand becomes effectively ordinal (reducesto Random Priority)

� CADA of Abdulkadiroglu, Che, and Yasuda : a nicecompromise between strategy-proofness and accounting forpreference intensity

� Boston mechanism: good efficiency properties in equilibrium(Abdulkadiroglu, Che, and Yasuda 2011) but failsstrategy-proofness even in the limit (Abdulkadiroglu andSonmez 2003, Kojima and Pathak 2009) and has badredistributive properties (Pathak and Sonmez 2008)

Trade-offs In Designing Cardinal Mechanisms

� Pseudo-market of Hylland and Zeckahuser 1979 :strategy-proof and efficient in the limit; challenge – how todecompose a random allocation?

� A-CEEI of Budish 2010: excellent with multi-unit demand,with single-unit demand becomes effectively ordinal (reducesto Random Priority)

� CADA of Abdulkadiroglu, Che, and Yasuda : a nicecompromise between strategy-proofness and accounting forpreference intensity

� Boston mechanism: good efficiency properties in equilibrium(Abdulkadiroglu, Che, and Yasuda 2011) but failsstrategy-proofness even in the limit (Abdulkadiroglu andSonmez 2003, Kojima and Pathak 2009) and has badredistributive properties (Pathak and Sonmez 2008)

Trade-offs In Designing Cardinal Mechanisms

� Pseudo-market of Hylland and Zeckahuser 1979 :strategy-proof and efficient in the limit; challenge – how todecompose a random allocation?

� A-CEEI of Budish 2010: excellent with multi-unit demand,with single-unit demand becomes effectively ordinal (reducesto Random Priority)

� CADA of Abdulkadiroglu, Che, and Yasuda : a nicecompromise between strategy-proofness and accounting forpreference intensity

� Boston mechanism: good efficiency properties in equilibrium(Abdulkadiroglu, Che, and Yasuda 2011) but failsstrategy-proofness even in the limit (Abdulkadiroglu andSonmez 2003, Kojima and Pathak 2009) and has badredistributive properties (Pathak and Sonmez 2008)

How to Decompose a Random Allocation?

Budish, Che, Kojima, Milgrom 2010 show how to decomposeallocations while preserving conjunctions of elementary constraintsof the form

c ≤�

(i ,h)∈C

P (i , h) ≤ c̄

and show how it matters in a variety of contextsCan we preserve other constraints?

c ≤�

(i ,h)∈C

P (i , h)−�

(i ,h)∈C �

P (i , h) ≤ c̄

Yes – Pycia and Unver (2010)

How to Decompose a Random Allocation?

Budish, Che, Kojima, Milgrom 2010 show how to decomposeallocations while preserving conjunctions of elementary constraintsof the form

c ≤�

(i ,h)∈C

P (i , h) ≤ c̄

and show how it matters in a variety of contextsCan we preserve other constraints?

c ≤�

(i ,h)∈C

P (i , h)−�

(i ,h)∈C �

P (i , h) ≤ c̄

Yes – Pycia and Unver (2010)

How to Decompose a Random Allocation?

Budish, Che, Kojima, Milgrom 2010 show how to decomposeallocations while preserving conjunctions of elementary constraintsof the form

c ≤�

(i ,h)∈C

P (i , h) ≤ c̄

and show how it matters in a variety of contextsCan we preserve other constraints?

c ≤�

(i ,h)∈C

P (i , h)−�

(i ,h)∈C �

P (i , h) ≤ c̄

Yes – Pycia and Unver (2010)

How to Decompose a Random Allocation?

Budish, Che, Kojima, Milgrom 2010 show how to decomposeallocations while preserving conjunctions of elementary constraintsof the form

c ≤�

(i ,h)∈C

P (i , h) ≤ c̄

and show how it matters in a variety of contextsCan we preserve other constraints?

c ≤�

(i ,h)∈C

P (i , h)−�

(i ,h)∈C �

P (i , h) ≤ c̄

Yes – Pycia and Unver (2010)

How to Decompose a Random Allocation?

Budish, Che, Kojima, Milgrom 2010 show how to decomposerandom allocations while preserving conjunctions of elementaryconstraints of the form

c ≤�

(i ,h)∈C

P (i , h) ≤ c̄

and show how it matters in a variety of contextsCan we preserve other constraints including constraints acrosstypes of agents’ profiles?

c ≤�

(θ,i ,h)∈C

P (θ) (i , h)−�

(θ,i ,h)∈C �

P (θ) (i , h) ≤ c̄

Yes – Pycia and Unver (2010)

How to Decompose a Random Allocation?

Budish, Che, Kojima, Milgrom 2010 show how to decomposerandom allocations while preserving conjunctions of elementaryconstraints of the form

c ≤�

(i ,h)∈C

P (i , h) ≤ c̄

and show how it matters in a variety of contextsCan we preserve other constraints including constraints acrosstypes of agents’ profiles?

c ≤�

(θ,i ,h)∈C

P (θ) (i , h)−�

(θ,i ,h)∈C �

P (θ) (i , h) ≤ c̄

Yes – Pycia and Unver (2010)